Remove padding from CL comparison operator

* Add support for processing left-over vector to comparison kernel.
* Combine native and quantized versions of CL comparison.

Resolves: COMPMID-6424
Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Change-Id: I31d43bdf0eab999cee6fa8144b5d8e921a1093e8
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10467
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

3 files changed, 68 insertions, 109 deletions

diff --git a/src/core/CL/cl_kernels/common/comparisons.cl b/src/core/CL/cl_kernels/common/comparisons.cl
index f05cb87835..00bb491f85 100644
--- a/src/core/CL/cl_kernels/common/comparisons.cl
+++ b/src/core/CL/cl_kernels/common/comparisons.cl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2021 Arm Limited.
+ * Copyright (c) 2018-2021, 2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -30,11 +30,13 @@
 #define LESS(x, y) ((x) < (y))
 #define LESSEQUAL(x, y) ((x) <= (y))
 
-#define DEFINE_KERNEL_STR(name) compare_##name
-#define DEFINE_KERNEL(name) DEFINE_KERNEL_STR(name)
+#ifdef IS_QUANTIZED
+#  define DEFINE_KERNEL_STR(name) compare_##name##_quantized
+#else // IS_QUANTIZED
+#  define DEFINE_KERNEL_STR(name) compare_##name
+#endif // IS_QUANTIZED
 
-#define DEFINE_KERNEL_QUANTIZED_STR(name) compare_##name##_quantized
-#define DEFINE_KERNEL_QUANTIZED(name) DEFINE_KERNEL_QUANTIZED_STR(name)
+#define DEFINE_KERNEL(name) DEFINE_KERNEL_STR(name)
 
 #if defined(DATA_TYPE) && defined(VEC_SIZE) && defined(OP) && defined(OP_NAME)
 /** This function compares two tensors.
@@ -73,78 +75,49 @@ __kernel void DEFINE_KERNEL(OP_NAME)(
     TENSOR3D_DECLARATION(in2),
     TENSOR3D_DECLARATION(out))
 {
-    // Get pixels pointer
-    Tensor3D in1 = CONVERT_TO_TENSOR3D_STRUCT(in1);
-    Tensor3D in2 = CONVERT_TO_TENSOR3D_STRUCT(in2);
-    Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(out);
+    int dst_x = max((int)get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE, 0);
 
-    // Load values
-    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
-    in_a = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in1.ptr);
-    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
-    in_b = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in2.ptr);
+#if VEC_SIZE_IN1 == 1
+    int in1_x = 0;
+#else // VEC_SIZE_IN1 == 1
+    int in1_x = dst_x;
+#endif // VEC_SIZE_IN1 == 1
 
-    // Calculate and store result
-    VSTORE(VEC_SIZE)
-    (CONVERT(OP(in_a, in_b), VEC_DATA_TYPE(uchar, VEC_SIZE)), 0, (__global uchar *)out.ptr);
-}
-#endif /* defined(DATA_TYPE) && defined(VEC_SIZE) && defined(OP) && defined(OP_NAME) */
+#if VEC_SIZE_IN2 == 1
+    int in2_x = 0;
+#else // VEC_SIZE_IN2 == 1
+    int in2_x = dst_x;
+#endif // VEC_SIZE_IN2 == 1
 
-#if defined(OFFSET_IN1) && defined(OFFSET_IN2) && defined(SCALE_IN1) && defined(SCALE_IN2)
-/** This function compares two quantized tensors.
- *
- * @note The inputs' data type need to be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=uchar
- * @note The quantization offset of the first operand must be passed at compile time using -DOFFSET_IN1, i.e. -DOFFSET_IN1=10
- * @note The quantization offset of the second operand must be passed at compile time using -DOFFSET_IN2, i.e. -DOFFSET_IN2=10
- * @note The quantization scale of the first operand must be passed at compile time using -DSCALE_IN1, i.e. -DSCALE_IN1=10
- * @note The quantization scale of the second operand must be passed at compile time using -DSCALE_IN2, i.e. -DSCALE_IN2=10
- *
- * @param[in]  in1_ptr                           Pointer to the source tensor. Supported data types: All quantized data types.
- * @param[in]  in1_stride_x                      Stride of the source tensor in X dimension (in bytes)
- * @param[in]  in1_step_x                        in1_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  in1_stride_y                      Stride of the source tensor in Y dimension (in bytes)
- * @param[in]  in1_step_y                        in1_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  in1_stride_z                      Stride of the source tensor in Z dimension (in bytes)
- * @param[in]  in1_step_z                        in1_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]  in1_offset_first_element_in_bytes The offset of the first element in the source tensor
- * @param[in]  in2_ptr                           Pointer to the source tensor. Supported data types: same as @p in1_ptr
- * @param[in]  in2_stride_x                      Stride of the source tensor in X dimension (in bytes)
- * @param[in]  in2_step_x                        in2_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  in2_stride_y                      Stride of the source tensor in Y dimension (in bytes)
- * @param[in]  in2_step_y                        in2_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  in2_stride_z                      Stride of the source tensor in Z dimension (in bytes)
- * @param[in]  in2_step_z                        in2_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]  in2_offset_first_element_in_bytes The offset of the first element in the source tensor
- * @param[out] out_ptr                           Pointer to the destination tensor. Supported data types: U8
- * @param[in]  out_stride_x                      Stride of the destination tensor in X dimension (in bytes)
- * @param[in]  out_step_x                        out_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  out_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
- * @param[in]  out_step_y                        out_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  out_stride_z                      Stride of the source tensor in Z dimension (in bytes)
- * @param[in]  out_step_z                        out_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]  out_offset_first_element_in_bytes The offset of the first element in the destination tensor
- */
-__kernel void DEFINE_KERNEL_QUANTIZED(OP_NAME)(
-    TENSOR3D_DECLARATION(in1),
-    TENSOR3D_DECLARATION(in2),
-    TENSOR3D_DECLARATION(out))
-{
-    // Get pixels pointer
-    Tensor3D in1 = CONVERT_TO_TENSOR3D_STRUCT(in1);
-    Tensor3D in2 = CONVERT_TO_TENSOR3D_STRUCT(in2);
-    Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(out);
+    int y = get_global_id(1);
+    int z = get_global_id(2);
 
-    int16 in_a = CONVERT(vload16(0, (__global DATA_TYPE *)in1.ptr), int16);
-    int16 in_b = CONVERT(vload16(0, (__global DATA_TYPE *)in2.ptr), int16);
+    in1_ptr += in1_offset_first_element_in_bytes + z * in1_stride_z + y * in1_stride_y + in1_x * sizeof(DATA_TYPE);
+    in2_ptr += in2_offset_first_element_in_bytes + z * in2_stride_z + y * in2_stride_y + in2_x * sizeof(DATA_TYPE);
+    out_ptr += out_offset_first_element_in_bytes + z * out_stride_z + y * out_stride_y + dst_x * sizeof(uchar);
 
-    in_a = in_a - (int16)((int)OFFSET_IN1);
-    in_b = in_b - (int16)((int)OFFSET_IN2);
+    // Load values
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) in_a = (VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))VLOAD(VEC_SIZE_IN1)(0, (__global DATA_TYPE *)in1_ptr);
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) in_b = (VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))VLOAD(VEC_SIZE_IN2)(0, (__global DATA_TYPE *)in2_ptr);
+
+    // Calculate and store result
+#ifdef IS_QUANTIZED
+    VEC_DATA_TYPE(int, VEC_SIZE) in_a_i32 = CONVERT(in_a, VEC_DATA_TYPE(int, VEC_SIZE));
+    VEC_DATA_TYPE(int, VEC_SIZE) in_b_i32 = CONVERT(in_b, VEC_DATA_TYPE(int, VEC_SIZE));
 
-    const float16 in1f32 = convert_float16(in_a) * (float16)((float)SCALE_IN1);
-    const float16 in2f32 = convert_float16(in_b) * (float16)((float)SCALE_IN2);
-    const int16   res    = OP(in1f32, in2f32);
+    VEC_DATA_TYPE(float, VEC_SIZE) in_a_fp = CONVERT(in_a_i32 - OFFSET_IN1, VEC_DATA_TYPE(float, VEC_SIZE)) * SCALE_IN1;
+    VEC_DATA_TYPE(float, VEC_SIZE) in_b_fp = CONVERT(in_b_i32 - OFFSET_IN2, VEC_DATA_TYPE(float, VEC_SIZE)) * SCALE_IN2;
+#else // IS_QUANTIZED
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) in_a_fp = in_a;
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) in_b_fp = in_b;
+#endif // IS_QUANTIZED
 
-    // Store result
-    vstore16(convert_uchar16(res), 0, (__global uchar *)out.ptr);
+#if VEC_SIZE == 1
+    uchar res0 = (uchar)select(0, 255, OP(in_a_fp, in_b_fp));
+#else // VEC_SIZE == 1
+    VEC_DATA_TYPE(uchar, VEC_SIZE) res0 = CONVERT(OP(in_a_fp, in_b_fp), VEC_DATA_TYPE(uchar, VEC_SIZE));
+#endif // VEC_SIZE == 1
+
+    STORE_VECTOR_SELECT(res, uchar, out_ptr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
 }
-#endif /* defined(OFFSET_IN1) && defined(OFFSET_IN2) && defined(SCALE_IN1) && defined(SCALE_IN2) */
\ No newline at end of file
+#endif /* defined(DATA_TYPE) && defined(VEC_SIZE) && defined(OP) && defined(OP_NAME) */
diff --git a/src/core/CL/kernels/CLComparisonKernel.cpp b/src/core/CL/kernels/CLComparisonKernel.cpp
index f27270733e..a0f9aca54a 100644
--- a/src/core/CL/kernels/CLComparisonKernel.cpp
+++ b/src/core/CL/kernels/CLComparisonKernel.cpp
@@ -25,6 +25,7 @@
 
 #include "arm_compute/core/CL/CLHelpers.h"
 #include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/utils/helpers/AdjustVecSize.h"
 #include "arm_compute/core/utils/StringUtils.h"
 
 #include "src/core/CL/CLValidate.h"
@@ -45,11 +46,6 @@ const std::map<ComparisonOperation, std::string> supported_comparison_ops = {
     {ComparisonOperation::Less, "LESS"},       {ComparisonOperation::LessEqual, "LESSEQUAL"},
 };
 
-int calculate_num_elems_processed_per_iteration(const ITensorInfo &input)
-{
-    return 16 / input.element_size();
-}
-
 Status validate_arguments(const ITensorInfo  &input1,
                           const ITensorInfo  &input2,
                           const ITensorInfo  &output,
@@ -77,26 +73,15 @@ Status validate_arguments(const ITensorInfo  &input1,
 std::pair<Status, Window> validate_and_configure_window(ITensorInfo &input1, ITensorInfo &input2, ITensorInfo &output)
 {
     const TensorShape &out_shape = TensorShape::broadcast_shape(input1.tensor_shape(), input2.tensor_shape());
-    const unsigned int num_elems_processed_per_iteration = calculate_num_elems_processed_per_iteration(input1);
+    const unsigned int num_elems_processed_per_iteration =
+        adjust_vec_size(16 / input1.element_size(), output.dimension(0));
 
     // Auto initialize output if not initialized
     auto_init_if_empty(output, out_shape, 1, DataType::U8, QuantizationInfo());
 
-    Window win        = calculate_max_window(out_shape, Steps(num_elems_processed_per_iteration));
-    Window win_input1 = win.broadcast_if_dimension_le_one(input1);
-    Window win_input2 = win.broadcast_if_dimension_le_one(input2);
-
-    AccessWindowHorizontal input1_access(&input1, 0, num_elems_processed_per_iteration);
-    AccessWindowHorizontal input2_access(&input2, 0, num_elems_processed_per_iteration);
-    AccessWindowHorizontal output_access(&output, 0, num_elems_processed_per_iteration);
-
-    bool window_changed = update_window_and_padding(win_input1, input1_access) ||
-                          update_window_and_padding(win_input2, input2_access) ||
-                          update_window_and_padding(win, output_access);
+    Window win = calculate_max_window(out_shape, Steps(num_elems_processed_per_iteration));
 
-    Status err =
-        (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
-    return std::make_pair(err, win);
+    return std::make_pair(Status{}, win);
 }
 } // namespace
 
@@ -133,11 +118,21 @@ void CLComparisonKernel::configure(const CLCompileContext &compile_context,
     const std::string &operation_name = supported_comparison_ops.at(operation);
     std::string        kernel_name    = "compare_" + lower_string(operation_name);
 
+    const unsigned int num_elems_processed_per_iteration =
+        adjust_vec_size(16 / input1->info()->element_size(), output->info()->dimension(0));
+
     // Set kernel build options
     std::set<std::string> build_opts;
     build_opts.emplace("-DDATA_TYPE=" + get_cl_type_from_data_type(input1->info()->data_type()));
-    build_opts.emplace("-DVEC_SIZE=" +
-                       support::cpp11::to_string(calculate_num_elems_processed_per_iteration(*input1->info())));
+    build_opts.emplace("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
+    build_opts.emplace("-DVEC_SIZE_LEFTOVER=" +
+                       support::cpp11::to_string(output->info()->dimension(0) % num_elems_processed_per_iteration));
+    build_opts.emplace(
+        "-DVEC_SIZE_IN1=" + //
+        support::cpp11::to_string(input1->info()->dimension(0) == 1 ? 1 : num_elems_processed_per_iteration));
+    build_opts.emplace(
+        "-DVEC_SIZE_IN2=" + //
+        support::cpp11::to_string(input2->info()->dimension(0) == 1 ? 1 : num_elems_processed_per_iteration));
     build_opts.emplace("-DOP=" + operation_name);
     build_opts.emplace("-DOP_NAME=" + lower_string(operation_name));
     if (is_data_type_quantized(input1->info()->data_type()))
@@ -145,6 +140,7 @@ void CLComparisonKernel::configure(const CLCompileContext &compile_context,
         const UniformQuantizationInfo iq1_info = input1->info()->quantization_info().uniform();
         const UniformQuantizationInfo iq2_info = input2->info()->quantization_info().uniform();
 
+        build_opts.emplace("-DIS_QUANTIZED");
         build_opts.emplace("-DOFFSET_IN1=" + support::cpp11::to_string(iq1_info.offset));
         build_opts.emplace("-DOFFSET_IN2=" + support::cpp11::to_string(iq2_info.offset));
         build_opts.emplace("-DSCALE_IN1=" + float_to_string_with_full_precision(iq1_info.scale));
@@ -228,13 +224,4 @@ void CLComparisonKernel::run(const Window &window, cl::CommandQueue &queue)
     } while (collapsed.slide_window_slice_3D(slice));
 }
 
-BorderSize CLComparisonKernel::border_size() const
-{
-    const int num_elems_processed_per_iteration = calculate_num_elems_processed_per_iteration(*_input1->info());
-
-    const unsigned int replicateSize =
-        _output->info()->dimension(0) - std::min(_input1->info()->dimension(0), _input2->info()->dimension(0));
-    const unsigned int border = std::min<unsigned int>(num_elems_processed_per_iteration - 1U, replicateSize);
-    return BorderSize{0, border, 0, 0};
-}
 } // namespace arm_compute
diff --git a/src/core/CL/kernels/CLComparisonKernel.h b/src/core/CL/kernels/CLComparisonKernel.h
index 174a6c9bf9..2fb4ba06b6 100644
--- a/src/core/CL/kernels/CLComparisonKernel.h
+++ b/src/core/CL/kernels/CLComparisonKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2020 Arm Limited.
+ * Copyright (c) 2018-2020, 2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -21,8 +21,8 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifndef ARM_COMPUTE_CLCOMPARISONKERNEL_H
-#define ARM_COMPUTE_CLCOMPARISONKERNEL_H
+#ifndef ACL_SRC_CORE_CL_KERNELS_CLCOMPARISONKERNEL_H
+#define ACL_SRC_CORE_CL_KERNELS_CLCOMPARISONKERNEL_H
 
 #include "arm_compute/core/Types.h"
 
@@ -85,8 +85,7 @@ public:
                            ComparisonOperation operation);
 
     // Inherited methods overridden:
-    void       run(const Window &window, cl::CommandQueue &queue) override;
-    BorderSize border_size() const override;
+    void run(const Window &window, cl::CommandQueue &queue) override;
 
 private:
     const ICLTensor *_input1; /**< Source tensor 1 */
@@ -94,4 +93,4 @@ private:
     ICLTensor       *_output; /**< Destination tensor */
 };
 } // namespace arm_compute
-#endif /* ARM_COMPUTE_CLCOMPARISONKERNEL_H */
+#endif // ACL_SRC_CORE_CL_KERNELS_CLCOMPARISONKERNEL_H